[CUDA] FPxINT quantized matmul for Hopper - #3160

Refs #2536.

This PR implements QMM for Hopper only, using CUTLASS.

Performance of FP16xINT4 with group size 64 on H100:

Independent C++ code for profiling the kernel

For small batch size (<=16) the performance data is similar to (though slower than) Machete so I think it is doing good. When M becomes larger than 32 it can no longer surpass cuBLAS, and it can be solved by increasing tile size, which I will fix in following PRs.

Another problem is this kernel behaves badly for small N/K (0.5x for n4096k4096), I'm not sure if it is something TMA kernels are not good at and we might have to fallback to Ampere optimized kernels.

There are also a few things need further optimizations:

1. The CUTLASS kernel requires scales/zeros to be N-major but MLX's quant format is K-major so there is a transpose involved for each call. CUDA graph can partly hide this overhead but I haven't profiled how much it hurts performance.
2. By shuffling the quantized weights we can use wider vectorized reads and faster int-to-float conversions, which would bring 17%~20% improvement. The 55_hopper_int4_bf16_gemm example implements a --shuffle switch that can be used to measure the performance difference.